How to Detect AI Referral Traffic in GA4 (UTM Setup for the 7 AI Engines)

AI engines send users to websites through one of four distinct routing paths. Each path lands the visitor in a different GA4 bucket, with different attribution accuracy. Understanding the four paths is the prerequisite for any meaningful measurement setup.

Path 1 — Tagged direct. The AI engine appends a recognisable parameter to the destination URL (most commonly utm_source set to the engine's domain). When the user clicks, GA4 sees the parameter and attributes the session to that source. This is the cleanest signal. ChatGPT routinely tags traffic this way; some other engines do too, inconsistently.

Path 2 — Untagged direct. The AI engine sends the user with no URL parameters and no referrer header. GA4 treats this as a Direct session — indistinguishable from a typed URL. AI traffic landing here is invisible without a custom signal. A non-trivial share of all AI clickthroughs lands in this bucket today.

Path 3 — Referrer-attributed. The user clicks from the engine's web interface and the browser sends a referrer header pointing to the engine's domain (for example, perplexity.ai, claude.ai, gemini.google.com). GA4 attributes the session to that domain as a Referral source. This is the second-cleanest signal after tagged direct.

Path 4 — Dark. The user reads the answer in the AI engine, copies a fact or link, opens a new browser session later, and types or searches for the brand independently. GA4 attributes this to Direct, Organic Search, or whatever channel the eventual click came from. The original AI exposure is invisible. This is the largest unmeasured channel and the reason every attribution model underestimates AI impact.

The seven canonical engines route traffic differently. Knowing each one's pattern lets you build attribution rules that capture what is capturable and acknowledge what is not.

ChatGPT (chatgpt.com). Routinely tags external links with utm_source set to chatgpt.com. The cleanest engine to measure. Most ChatGPT-referred sessions land in GA4 with the source parameter intact and attribute correctly.

Claude (claude.ai). Less consistent. Some clickthroughs pass a referrer of claude.ai, attributing as Referral; others arrive with no referrer and land in Direct. Tagging is uncommon. Mixed-signal engine.

Perplexity (perplexity.ai). Strong referrer-attribution pattern. Most Perplexity clickthroughs arrive with the referrer header pointing to perplexity.ai. GA4 attributes them as Referral. Easy to spot in standard reports under Referral sources.

Gemini (gemini.google.com). Variable. Sometimes passes a Google-domain referrer; sometimes arrives with no referrer. A meaningful share of Gemini traffic lands in Direct or in the broader Google Organic bucket and is hard to attribute precisely.

Google AI Overviews. Inline within Google search results, so any clickthrough from within the AI Overview block typically attributes as Google Organic — GA4 does not distinguish the AI Overview click from a traditional ranked-result click. Attribution to AI Overviews specifically requires more advanced setup (typically GSC API plus custom event tracking).

Google AI Mode. Similar to AI Overviews. Clickthroughs from the AI-first interface generally attribute as Google Organic. Distinguishing AI Mode traffic from standard Google Organic is currently difficult without GSC-side instrumentation.

Microsoft Copilot (copilot.microsoft.com). Inconsistent. Some Copilot sessions pass copilot.microsoft.com as referrer; some arrive with no referrer; some pass bing.com as referrer. Mixed Direct, Referral, and Bing Organic attribution.

Before adding any custom configuration, audit what GA4 already shows about AI-referred sessions. Often more is already there than the brand realises — but it is buried in default reports that group it ambiguously.

Open the standard Acquisition reports in GA4 and apply the following filters in sequence. First, in the Traffic Acquisition report, filter the Session source dimension to show only entries containing chatgpt, claude, perplexity, gemini, or copilot. Any rows that appear here are already-attributed AI sessions — captured via the tagged-direct or referrer-attributed paths described in the previous section.

Second, in the same report, look at the rows for Direct traffic and Google Organic. AI sessions are mixed into both of these buckets via the untagged-direct and referrer-mixed paths. There is no clean way to extract them from these aggregates without custom configuration, but the proportional growth of Direct and Google Organic over time often hides a meaningful AI-attribution share.

Third, check the Referrals report (Acquisition then Traffic Acquisition, filtered to Referral). Look for entries from perplexity.ai, claude.ai, copilot.microsoft.com, gemini.google.com, and similar domains. These are the referrer-attributed sessions and represent the cleanest measurable AI signal after tagged direct.

The audit will typically reveal that the brand already has measurable AI traffic — usually somewhere between half a percent and three percent of total sessions for an Irish SME with active AEO investment. That baseline is the starting point against which all the subsequent setup work compounds.

The default GA4 setup is not optimised for AI attribution. Two configuration changes meaningfully improve what gets captured.

Custom dimension for AI Source. Add a custom session-scoped dimension called something like "AI Engine Source". The tag management rule populating it should look at the document referrer plus the URL search parameters, and write a value when the referrer or source matches one of the seven canonical AI engine domains. Treat any match (regardless of which path the engine used to route the traffic) as AI-attributed, with the engine name as the dimension value.

This custom dimension is the load-bearing piece. It unifies the four routing paths into one explicit AI attribution field. Once populated, every report can be sliced by AI Source as cleanly as by Channel or Source.

Channel grouping override. The default GA4 channel grouping does not include an AI Search channel. Add a custom channel grouping (in GA4 Admin under Data display, then Channel groups) that promotes any session with the AI Source dimension populated into a dedicated "AI Search" channel. This makes AI sessions visible in the standard Acquisition Overview report alongside Organic Search, Direct, Paid, and the other default channels.

Both changes are configuration only — no developer work beyond a small tag management rule. A competent GA4 administrator can implement both in under two hours. The data starts accumulating from the moment the rule is live; it does not backfill.

Once the custom dimension is live, build a dedicated GA4 Exploration that surfaces AI-attributed sessions and their downstream behaviour. The Exploration is what the brand uses to monitor AI traffic ongoing, and to justify continued AEO investment to stakeholders.

The recommended report structure includes:

• Time series row. Sessions over time, segmented by AI Engine Source value. Shows growth or decline per engine.
• AI engine summary table. Rows are AI Engine Source values; columns are Sessions, Engaged Sessions, Average Engagement Time, Conversions. Surfaces which AI engines drive the highest-intent traffic.
• Conversion path table. Use the Path Exploration template, with the starting node set to an AI-attributed first session. Shows how AI-discovered users behave across subsequent sessions and whether they eventually convert.
• Geo overlay. Filter the Exploration to Ireland-located sessions only. For an Irish SME, AI traffic from outside Ireland is usually less commercially relevant and dilutes the signal in aggregate reports.

Share the Exploration with anyone reviewing AI visibility outcomes — clients, stakeholders, internal management. The single most valuable insight tends to be the comparison of AI-attributed conversion rate against Direct and Organic conversion rates. AI-attributed sessions typically convert at a higher rate than typed-URL Direct sessions because the user arrived with a recommendation already loaded in their head.

Three honest caveats every brand needs to acknowledge when reporting AI-attribution data. Ignoring them produces over-claimed numbers that collapse under scrutiny; acknowledging them makes the rest of the data credible.

Dark traffic is large. The fourth routing path described above — the user reads an AI answer, copies a fact or remembers the brand, returns later via a different channel — is invisible to GA4 by design. There is no measurement method that captures it directly. Estimates of dark AI traffic range from 30 to 60 per cent of total AI-influenced sessions, based on the gap between AI-citation impression volume (visible in tools like SurgeGraph or ZeroRank) and AI-attributed session volume in GA4. Always present GA4 AI-attribution numbers as a lower bound, not a total.

Referrer stripping is widespread. Modern browsers ship with referrer policies that strip or shorten the Referer header on outbound clicks. Privacy-focused browsers (Brave, Safari with strict settings, Firefox with enhanced tracking protection) strip more aggressively than Chrome. The result: the referrer-attributed path captures fewer sessions than the engine actually sends. A 10 to 30 per cent under-count from this alone is typical.

Sampling and attribution model effects. GA4 applies sampling on high-volume queries and uses a data-driven attribution model that distributes conversion credit across touchpoints. AI-attributed sessions that are the first touch in a multi-touch journey may receive only a fraction of the credit even when they were the originating discovery channel. Use the First Touch attribution model when explicitly measuring AI discovery, not the default Data-Driven model.

The honest summary: GA4 AI attribution is directionally correct but quantitatively under-counted. The numbers are useful for measuring growth and comparing engines; they are unreliable as a precise revenue-attribution input. Pair them with citation-volume measurements from SurgeGraph or ZeroRank for a complete picture.